UCSC-VLAA/CIK-Bench
收藏资源简介:
CIK-Bench是一个用于评估OpenClaw个人AI代理安全性的数据集,针对持久状态中毒攻击。它基于CIK分类法,将攻击分为能力(Capability)、身份(Identity)和知识(Knowledge)三个维度。数据集包含88个攻击案例,涵盖12个影响场景,分为隐私泄露(Privacy Leakage)和风险不可逆操作(Risky Irreversible Operations)两大类别,每个类别下还有子类别。此外,数据集还包含一组匹配的良性案例用于防御评估。数据集以结构化行和原始模板树两种形式提供,支持通过HuggingFace的datasets库加载和使用。
CIK-Bench is a dataset designed to evaluate the safety of the OpenClaw personal AI agent against persistent-state poisoning attacks. It implements the CIK taxonomy, organizing attacks into three dimensions: Capability, Identity, and Knowledge. The dataset includes 88 attack cases across 12 impact scenarios, categorized into Privacy Leakage and Risky Irreversible Operations, each with subcategories. It also includes a matched set of benign cases for defense evaluation. The dataset is provided in two forms: structured rows and a raw template tree, and can be loaded and used via the HuggingFace datasets library.




